Uncertainty of Spatial Segmentation of High-Speed Videoendoscopy and Its Temporal and Spatial Dependency.

Abstract

Objective: Spatial segmentation of high-speed videoendoscopy (HSV) is the process that detects the edges of the vocal folds and represents them in analytic form. The level of spatial segmentation uncertainty (ie, how close vs. far apart different experts marked the edges of the vocal folds) can have a great impact on the level of uncertainty of the final measures (ie, their dispersion). This study quantified the uncertainty of spatial segmentation and investigated its dependency on the phase of the glottal cycle and the location of vocal fold edges along the anterior-posterior direction.

Method: Three experts manually segmented the vocal fold edges of twelve HSV recordings using an iterative process consisting of an initial segmentation followed by a blinded reconciliation phase. Segmentation uncertainty was computed as the distance in pixels between the three-segmented edges at the end of the iterative process. The relationships between segmentation uncertainty and different sections of the glottis along the anterior-posterior direction and the relationships between segmentation uncertainty and different phases of the glottal cycle were quantified.

Results: Segmentation uncertainties of the anterior and the posterior sections of the glottis were significantly higher than the middle section, while uncertainty of the anterior section was the highest and 40% larger than the middle section. The average segmentation uncertainty and normalized glottal area were positively correlated. Segmentation uncertainty of the most open glottal configurations was 31% larger than the most closed glottal configuration.

Conclusion: The uncertainty of spatial segmentation of the vocal fold edges depends on the phase of the glottal cycle and the location of the edge along the anterior-posterior direction; hence, it is expected for different HSV measures to have different levels of uncertainties. The implications of these findings for vocal fold velocity measures are discussed. Additionally, the findings from this study could provide direction for future automated spatial segmentation methods and for creating a robust and reliable automated HSV processing pipeline.